tty: throttling race fix

The tty throttling code can race due to the lock drops. It takes very high
loads but this has been observed and verified by Rob Duncan.

The basic problem is that on an SMP box we can go

	CPU #1				CPU #2
	need to throttle ?
	suppose we should		buffer space cleared
					are we throttled
					yes ? - unthrottle
	call throttle method

This changeet take the termios lock to protect against this. The termios
lock isn't the initial obvious candidate but many implementations of throttle
methods already need to poke around their own termios structures (and nobody
really locks them against a racing change of flow control).

This does mean that anyone who is setting tty->low_latency = 1 and then
calling tty_flip_buffer_push from their unthrottle method is going to end up
collapsing in a pile of locks. However we've removed all the known bogus
users of low_latency = 1 and such use isn't safe anyway for other reasons so
catching it would be an improvement.

Signed-off-by: Alan Cox <alan@lxorguk.ukuu.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 4 insertions, 2 deletions

diff --git a/include/linux/tty_driver.h b/include/linux/tty_driver.h
index bcba84ea2d86..3566129384a4 100644
--- a/include/linux/tty_driver.h
+++ b/include/linux/tty_driver.h
@@ -127,7 +127,8 @@
  *      the line discipline are close to full, and it should somehow
  *      signal that no more characters should be sent to the tty.
  *
- *      Optional: Always invoke via tty_throttle();
+ *      Optional: Always invoke via tty_throttle(), called under the
+ *      termios lock.
  * 
  * void (*unthrottle)(struct tty_struct * tty);
  *
@@ -135,7 +136,8 @@
  *      that characters can now be sent to the tty without fear of
  *      overrunning the input buffers of the line disciplines.
  * 
- *      Optional: Always invoke via tty_unthrottle();
+ *      Optional: Always invoke via tty_unthrottle(), called under the
+ *      termios lock.
  *
  * void (*stop)(struct tty_struct *tty);
  *